_PyPegen_is_memoized() has a complexity of O(n)

Using Linux perf, I noticed that the Python parser spends a significant time in the _PyPegen_is_memoized() function which iterates on a linked list to find a value:

int  // bool
_PyPegen_is_memoized(Parser *p, int type, void *pres)
{
    if (p->mark == p->fill) {
        if (_PyPegen_fill_token(p) < 0) {
            p->error_indicator = 1;
            return -1;
        }
    }

    Token *t = p->tokens[p->mark];

    for (Memo *m = t->memo; m != NULL; m = m->next) {
        if (m->type == type) {
#if defined(PY_DEBUG)
            if (0 <= type && type < NSTATISTICS) {
                long count = m->mark - p->mark;
                // A memoized negative result counts for one.
                if (count <= 0) {
                    count = 1;
                }
                memo_statistics[type] += count;
            }
#endif
            p->mark = m->mark;
            *(void **)(pres) = m->node;
            return 1;
        }
    }
    return 0;
}

script.py:

import tokenize

# wc -l $(find -name "*.py")|sort -n
large_files = """
    5259 ./Tools/clinic/clinic.py
    5553 ./Lib/test/test_email/test_email.py
    5577 ./Lib/test/test_argparse.py
    5659 ./Lib/test/test_logging.py
    5806 ./Lib/test/test_descr.py
    5878 ./Lib/test/test_decimal.py
    5993 ./Lib/test/_test_multiprocessing.py
    6425 ./Lib/_pydecimal.py
    6626 ./Lib/test/datetimetester.py
    6664 ./Lib/test/test_socket.py
    7325 ./Lib/test/test_typing.py
   15534 ./Lib/pydoc_data/topics.py
"""
large_files = [line.split()[-1] for line in large_files.splitlines() if line.strip()]

files = []
for filename in large_files:
    with tokenize.open(filename) as fp:
        content = fp.read()
    files.append((filename, content))

for loops in range(5):
    for filename, content in files:
        print(filename)
        compile(content, filename, "exec")

Linux perf says that overall, Python spent 7% of its runtime in this function:

$ perf record ./python script.py
$ perf report
Samples: 10K of event 'cycles:u', Event count (approx.): 7308792716
Overhead  Command  Shared Object         Symbol
   7,35%  python   python                [.] _PyPegen_is_memoized
   5,18%  python   python                [.] assemble
   3,80%  python   python                [.] _PyPegen_expect_token
   3,53%  python   python                [.] unicodekeys_lookup_unicode
   3,06%  python   python                [.] tok_get
   2,59%  python   python                [.] _PyPegen_update_memo
   2,27%  python   python                [.] _Py_dict_lookup
   2,19%  python   python                [.] _PyObject_Free
(...)

@pablogsal

@pablogsal @lysnikolaou @isidentical: Would it be possible to maintain a hashtable to make this lookup function faster?

If you consider that it's not worth it, just you can close my issue ;-)

The benchmark comes from: #93106

@pablogsal

@pablogsal @lysnikolaou @isidentical: Would it be possible to maintain a hashtable to make this lookup function faster?

If you consider that it's not worth it, just you can close my issue ;-)

We already tried to use a hash table, a btree, a static array, a dynamic array and a skip list and all of them were slower. If you manage to get a hash table to perform faster, I'm happy to merge it :)

I tried yet again to use the hash table from pycore_hashtable.h and is a performance disaster. I took also some statistics and the linked lists average number of elements is between 1 and 10 values so a hash table is an overkill, either per token or one in the parser. Also I checked and the average percentage of times the function returns true (so a value is found) is around 75%, which means that adding a bloom filter or similar in front is also going to slow down the thing quite a lot.

I'm closing the issue, but if you manage to get a hash table or other structure running and is faster, please feel free to reopen.

@isidentical

@isidentical @lysnikolaou If you want to give it a go, please feel free to try :)

I tried yet again to use the hash table from pycore_hashtable.h and is a performance disaster.

That's what I would try. I trust you that it's worse ;-) Ok, good that you already knew about it!

Navigation Menu

Repository navigation

_PyPegen_is_memoized() has a complexity of O(n) #93289

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

_PyPegen_is_memoized() has a complexity of O(n) #93289

Description

Activity

vstinner commented on May 27, 2022

vstinner commented on May 27, 2022

pablogsal commented on May 27, 2022

pablogsal commented on May 28, 2022

pablogsal commented on May 28, 2022

vstinner commented on May 28, 2022

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions